Middle DropBlock network for enhanced activation in person re-identification

In person re-identification tasks, especially in non-overlapping camera systems, all instances matching a query image must be accurately retrieved from a large image collection. To address the susceptibility of models to background interference and low-information regions, a Middle DropBlock Network is proposed. This network guides the model to focus on foreground areas, thereby enhancing feature discriminability. The network architecture consists of three streams: (1) the Global Stream, which extracts rich image information from the backbone network; (2) the Middle DropBlock Stream, which encourages the backbone network to encode low-information regions to enhance feature distinctiveness; and (3) the Regularizer Stream, which mitigates noise introduced during feature dropout. Extensive experiments conducted on challenging datasets demonstrate that this method outperforms existing approaches in accurate identification and feature extraction. The results further indicate that the Middle DropBlock Network produces clearer activation maps, focusing the model’s attention on key informative regions of input images, thus significantly improving person re-identification performance.